Sie sind auf Seite 1von 47

Introduction

What is a wavelet
Wavelet for images
Wavelets & PDEs

Wavelet Methods for the solution of PDEs:


recent advances and future developments
Silvia Bertoluzza1
1 Istituto

di Matematica applicata e Tecnologie Informatiche del CNR, Pavia

Vienna, December 12th, 2006

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Layout

Layout of the talk

what is a wavelet ?
wavelets for image compression
wavelets for PDEs
adaptivity
nonlinear wavelet methods for PDEs
wavelet stabilization methods

the problem of the geometry


domain decompositiom
wavelets for FEM

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Example: The Haar Basis


Vj : p.w. const. on uniform grid, step h = 2j
Basis: translated of

Projection on coarser space Vj1


Basis: translated of

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

What is the difference


6

Basis: translated of

Difference
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Example: The P1 Interpolating Wavelets


Vj : p.w. linears on uniform grid, step h = 2j
6
B B B B B B B B 
B B B B B B B B
B B B B B B B B
B B B B B B B B
B
 B B B B B B B -

coarser space Vj1

Basis: translated of
B
 B



B
B

Basis: translated of

6
J

J
J
J
J

J
J
J
J

J
J
J
J
J

Silvia Bertoluzza

J
J

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

What is the difference


6
J B

J B

J B

J B

J B

J B

J B

J B

B
B
B
J
J
J


J
B

J B 
J B 
J B 
J B

J B 
J B 
J B 
J B
JJBB

JJBB

JJBB

JJBB


Difference
6
B
 B





B
 B
B


B


B 
BB

B
 B
B
B



B 
BB

B
 B
B


B


B 
BB
Silvia Bertoluzza

B
B
B
BB

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

In summary:
Space on the grid k/2j
6

6
B B B B B B B B 
B B B B B B B B
B B B B B B B B
B B B B B B B B
B
 B B B B B B B -

bases for Vj by dilation and translation of


B
 B



Silvia Bertoluzza

B
B
B

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

inclusion: coarser space finer space


difference spaces Wj : Vj+1 = Vj Wj
6

B 
B

B 
B




B
 B

B
 B
 B

B
 B

B
 B


B 
B

B
B
B
-

Bases for Wj by dilation and translation of


B
 B
-




Silvia Bertoluzza

B
B
B

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Vj
HH
HH

Vj1

Vj = Vj1 Wj1

Wj1

H
HH
H

Vj2

Vj = Vj2 Wj1 Wj1

Wj2

..
.
H
HH
H

V0

Vj = V0 j1
m=0 Wm

W0

L2 = V0
m=0 Wm
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Wavelet basis ' the whole structure:


Vj Vj+1
Wj : Vj+1 = Vj Wj
functions and generating bases for Vj and Wj by dilations
and translation.

Classes of wavelets
(Infinitely) many examples in which this structure is replicated
classes of wavelet bases

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Spline

Meyer

Daubechies

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Space-Frequency localization

Fundamental property
The mother wavelet is well localized both in space and in
frequency
(Recall: a function f and its Fourier transform f cannot be both compactly supported)

Consequences:
Characterization of function spaces
Singularity detection

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Characterization of Functional Spaces


wavelet function
dilated and translated versions:
jk = (2j x k)2j/2s
Basis for L2 (R)H s (R)
L2 H s (R) = linear span of {jk , (j, k) Z}
Rieszs basis property:
k

cjk jk k2L2 (R)H s (R) '

jk

Silvia Bertoluzza

|cjk |2

jk

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Example: Haar
Example: The P1 Interpolating Wavelets
Wavelet structure
Properties of wavelets

Besov spaces

s (R)
Basis for Bq,q
s
Bq,q
(R) = linear span of {jk , (j, k) Z}

Rieszs basis property for


k

cjk jk k

jk

s (R)
Bq,q

'

|cjk |q

jk

Remark
s plays a crucial role in non linear wavelet
The Besov space Bq,q
approximation
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

First wavelet application:Image compression

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

RGB Image:
n m pixels
3 numbers per pixel: rnk = red, gnk = green, bnk = blue
3 functions:
r=

rnk jnk ,

g=

Silvia Bertoluzza

gnk jnk ,

b=

bnk jnk

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Compression
each colour component decomposed in wavelet basis

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Compression

each wavelet coefficient is quantised:


high frequency less important to the human eye
small coefficients are set to 0
coefficients corresponding to higher frequencies represented
with less bits
lots of zeros where image is smooth!

resulting quantised image is encoded (using zip type


algorithm)

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Results - CR: 50:1

Original size = 50 compressed size


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Results - CR: 200:1

Original size = 200 compressed size


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Comparison with JPEG - C.R. 200:1

Original size = 200 compressed size


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Digital images
Compression
Comparison with JPEG

Remark
The standard JPEG2000 is based on wavelets

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelets and PDEs


late 80s: application of wavelets in image and signal
processing (image compression, seismic signal analysis, edge
detection, denoising)
early 90
initially straightforward application
first exploitations of the potential of wavelets for adaptivity,
preconditioning
application fields limited by intrinsic difficulties (geometry,
computation of fundamental quantities) [(89: compactly
supported wavelets,93: wavelets on the interval)]
[Glowinski et al, Maday et al., Liandrat et al, Dahmen et al.]
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Adaptivity

isolated singularity in a regular background


coeffs of wavelets not seeing the singularities are small
corresponding functions not needed
pb: position of singularities not known
size of coeffs information on singularities simple
refining and derefining strategy

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

solve on a finite dimensional subspace look at solution


detect singularities adapt the finite dimensional subspace
start again

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Linear vs. Nonlinear approximation

Linear approximation: fix a finite dimensional subspace Vh and


look for an approximation of f in Vh
piecewise linears on a fixed triangulation
polynomials of degree n
trigonometric polynomials of degree n
wavelet type Vj space

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Nonlinear approximation: subspace Vh for approximating f is


chosen depending on f
piecewise linears on a free triangulation
h p approximations
linear combination of n wavelet functions
(Other examples of nonlinear approximation: rational polynomials)

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Best N-terms approximation

Space V (Ex: L2 ), basis {n , n Z} verifying


X
X
k
cn n k2V =
cn2 = k~c k`2
n

Problem
Given f find the best approximation of the form
f

N
X

nk nk

k=1

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

How do we do it?

Error E
kE k2V =

N
X

|nk cnk |2 +

k=1

|cnk |2

n6=nk

Minimise the error:


nk cnk = 0
P
2
n6=nk |cnk | as small as possible

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Define a nonlinear projector

function to approximate f =

n cn n

reordering of the coefficients:


|cn0 | |cn1 | |cn1 | |cnk |

truncate the sum by retaining the N biggest coefficients


PN (f ) =

N
X

cnk nk

k=0

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Theorem.
If the sequence of coeffs ~c is in `q with q < 2 then
kE kV . N (1/q+1/2)
[DeVore]

Remark
If only kf kV ' k~c k`2 then PN (f ) is only a nearly best
approximation

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Nonlinear solution of PDEs

Problem
Find the (nearly-)best N-terms approximation to the solution u of
a PDE
Au = f
(A nice differential operator (symmetric, positive definite), ex:
A = )

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Step 1. rewrite the PDE as a problem in `2 on the sequence ~u of


wavelet coefficients of u
A~u = ~f

Step 2. fixed point formulation


~u = ~u (~f A~u )

Step 3. iterative scheme


~u n = ~u n1 (~f A~u n1 )

No approximation until now


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Richardson Scheme

initial guess ~u 0 = 0
~u n ~u n+1
preview a set n+1 containing all the indexes which will not be
discarded by the non linear projector
compute rnn = ~fn (
A~u n )n

for n n+1

~u n+1 = PN (~u n + rn )

wavelet properties = Richardson scheme converges


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Theorem
For 1 0 it holds:
stability: we have
k~u n k`2 . k~f k`2 + k~u 0 k`2 , n N.
approximation error estimate: for ~e n = ~u n ~u it holds:
k~e n k`2 n k~eo k`2 +

1
1
C
N ( + 2 ) ,
1

where C is a constant depending only on the initial data, and


<1
[Cohen, Dahmen, De Vore], [S.B. & Verani]
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelet Stabilization

V Hilbert space
V 0 its dual
A : V V 0 positive, semi-definite and boundedly invertible

Problem
Find numerical solution to
Au = f

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Observation
If A was positive definite Galerkin projection on any finite
dimensional subspace is optimal

Idea
positive definite and f V 0 such that
Find A
is positive definite
A
= f
the solution of Au = f coincides with the solution to Au

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Recipe
[Baiocchi-Brezzi,S.B.]

Let (, )V 0 denote the scalar product in V 0


defined by
The linear form A
v >=< Au, v > +(Au, Av )V 0
< Au,
is positive definite!
If Au = f then
v >=< f , v > +(f , Av )V 0 =:< f, v >
< Au,

Problem
How do we evaluate (, )V 0 in practice?
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelet realization of scalar product


[S.B.-Kunoth]

B = i basis for V 0 (or for any subspace H dense in V 0


norm equivalence of the form
X
X
k
ci i k2V 0 '
|ci |2
i

set
[

X
i

ci i ,

X
i

di i ]V 0 =

ci di

[, ]V 0 is an equivalent scalar product for V 0


Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelets: Pros

space-frequency localisation norm equivalences and


characterization of singularities
preconditioning
adaptivity
nonlinear approximation
easy computable equivalent norms and scalar products
applied to many different problems (particularly good results
for the simulation of turbulence)

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelets: Cons

potential of wavelet limited by serious drawbacks


constants in the inequalities are sometimes big
wavelets now known in closed form computation of
fundamental quantities is complicated
geometry

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Geometry

wavelets initially defined on R


Rd
]0, 1[
]0, 1[d

real-life application are much more complicated

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

Silvia Bertoluzza

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Option # 1. Domain decomposition


Domain split in (square) patches, basis for the domain is
constructed out of the bases for the patches
conforming domain decomposition:
strong regularity across the interfaces between patches for the
functions spanning the Vj
difficulty: continuity of the wavelet function
solution is quite technical
very small patches are not feasible
[Canuto, Tabacco, Urban - Dahmen, Kunoth, Urban - Cohen,
Masson]
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

non conforming domain decomposition


weak continuity across the interfaces: the jump must be
orthogonal to a suitable space
allows to use different type of functions in different patches
FE (piecewise polynomial) where needed (strange geometries,
small patches)
wavelets where possible (squared, sufficiently big patches)

[S.B., Falletta, Piquemal]

Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Option # 2: wavelet type FE

Family of triangulations Tj w. triangles of diameter h = 2j


Tj obtained from Tj1 by refinement: each triangle is cut in
four triangles
space Vj of continuous functions, linear on each triangle
Vj Vj+1
Pb: How to define Wj
Vj+1 = Vj Wj
the fundamental properties are replicated
the functions are practical to use
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Introduction
What is a wavelet
Wavelet for images
Wavelets & PDEs

History
Nonlinear wavelet methods
Wavelet stabilization
Open issues

Conclusions
wavelets powerful tool for representing and manipulating
different kind of data
cheap representation of solution of PDEs adaptivity
strong limitation to the applicability to real life problems
coupling with classical methods allows overcoming some of
the limitations

Remark
Replicating the fundamental properties of wavelets in the
framework of classical methods could lead to new and powerful
algorithms
Silvia Bertoluzza

Wavelet Methods for the solution of PDEs: recent advances and

Das könnte Ihnen auch gefallen